提交voiceprint代码

src/voiceprint/create_data.py

@@ -0,0 +1,108 @@
+import os
+import random
+import sys
+
+import librosa
+import numpy as np
+import tensorflow as tf
+from tqdm import tqdm
+
+
+# 获取浮点数组
+def _float_feature(value):
+    if not isinstance(value, list):
+        value = [value]
+    return tf.train.Feature(float_list=tf.train.FloatList(value=value))
+
+
+# 获取整型数据
+def _int64_feature(value):
+    if not isinstance(value, list):
+        value = [value]
+    return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
+
+
+# 把数据添加到TFRecord中
+def data_example(data, label):
+    feature = {
+        'data': _float_feature(data),
+        'label': _int64_feature(label),
+    }
+    return tf.train.Example(features=tf.train.Features(feature=feature))
+
+
+# 开始创建tfrecord数据
+def create_data_tfrecord(data_list_path, save_path):
+    with open(data_list_path, 'r') as f:
+        data = f.readlines()
+    with tf.io.TFRecordWriter(save_path) as writer:
+        for d in data:
+            try:
+                path, label = d.replace('\n', '').split('\t')
+                wav, sr = librosa.load(path, sr=16000)
+                intervals = librosa.effects.split(wav, top_db=20)
+                wav_output = []
+                # [可能需要修改参数] 音频长度 16000 * 秒数
+                wav_len = int(16000 * 2.04)
+                for sliced in intervals:
+                    wav_output.extend(wav[sliced[0]:sliced[1]])
+                for i in range(20):
+                    # 裁剪过长的音频，过短的补0
+                    if len(wav_output) > wav_len:
+                        l = len(wav_output) - wav_len
+                        r = random.randint(0, l)
+                        wav_output = wav_output[r:wav_len + r]
+                    else:
+                        wav_output.extend(np.zeros(shape=[wav_len - len(wav_output)], dtype=np.float32))
+                    wav_output = np.array(wav_output)
+                    # 转成梅尔频谱
+                    ps = librosa.feature.melspectrogram(y=wav_output, sr=sr, hop_length=256).reshape(-1).tolist()
+                    # [可能需要修改参数] 梅尔频谱shape ，librosa.feature.melspectrogram(y=wav_output, sr=sr, hop_length=256).shape
+                    if len(ps) != 128 * 128: continue
+                    tf_example = data_example(ps, int(label))
+                    writer.write(tf_example.SerializeToString())
+                    if len(wav_output) <= wav_len:
+                        break
+            except Exception as e:
+                print(e)
+
+
+# 生成数据列表
+def get_data_list(audio_path, list_path):
+    files = os.listdir(audio_path)
+
+    f_train = open(os.path.join(list_path, 'train_list.txt'), 'w')
+    f_test = open(os.path.join(list_path, 'test_list.txt'), 'w')
+
+    sound_sum = 0
+    s = set()
+    label = {}
+    for file in files:
+        if '.wav' not in file:
+            continue
+        index = file.find('-')
+        name = file[0:index]  # 取出每个wav文件的名字
+        print(name)
+        if name not in s:
+            label[name] = len(s)
+            s.add(name)
+        sound_path = os.path.join(audio_path, file)
+        if sound_sum % 10 == 0:  # 从第0个文件开始放入test，顺序为10的倍数的文件放入test
+            f_test.write('%s\t%d\n' % (sound_path.replace('\\', '/'), label[name]))
+        else:
+            f_train.write('%s\t%d\n' % (sound_path.replace('\\', '/'), label[name]))
+        sound_sum += 1
+
+    f_test.close()
+    f_train.close()
+
+
+if __name__ == '__main__':
+    trainername = sys.argv[1]
+    print('trainername = ' + trainername)
+    print(sys.path[0])
+    get_data_list(sys.path[0] + '/' + 'dataset/audio' + '/' + trainername, sys.path[0] + '/' + 'dataset')
+    create_data_tfrecord(sys.path[0] + '/' + 'dataset/train_list.txt', sys.path[0] + '/' + 'dataset/train.tfrecord')
+    create_data_tfrecord(sys.path[0] + '/' + 'dataset/test_list.txt',  sys.path[0] + '/' + 'dataset/test.tfrecord')
+    command = "python " + sys.path[0] + "/" + "train.py" + " " + trainername
+    os.system(command)

src/voiceprint/infer_contrast.py

@@ -0,0 +1,101 @@
+import os
+import random
+
+import librosa
+import numpy as np
+import tensorflow as tf
+from tensorflow.keras.models import Model
+
+
+layer_name = 'global_max_pooling2d'
+#model = tf.keras.models.load_model('models/resnet.h5')
+#intermediate_layer_model = Model(inputs=model.input, outputs=model.get_layer(layer_name).output)
+
+
+# 读取音频数据
+def load_data(data_path):
+    wav, sr = librosa.load(data_path, sr=16000)
+    intervals = librosa.effects.split(wav, top_db=20)
+    wav_output = []
+    for sliced in intervals:
+        wav_output.extend(wav[sliced[0]:sliced[1]])
+    assert len(wav_output) >= 8000, "有效音频小于0.5s"
+    wav_output = np.array(wav_output)
+    ps = librosa.feature.melspectrogram(y=wav_output, sr=sr, hop_length=256).astype(np.float32)
+    ps = ps[np.newaxis, ..., np.newaxis]
+    return ps
+
+
+def infer(audio_path):
+    data = load_data(audio_path)
+    feature = intermediate_layer_model.predict(data)
+    return feature
+
+
+if __name__ == '__main__':
+    # 要预测的两个人的音频文件
+    # person1 = 'dataset/wc-5.wav'
+    # person2 = 'dataset/wc-1.wav'
+    # feature1 = infer(person1)[0]
+    # feature2 = infer(person2)[0]
+    # # 对角余弦值
+    # dist = np.dot(feature1, feature2) / (np.linalg.norm(feature1) * np.linalg.norm(feature2))
+    # if dist > 0.92:
+    #     print("%s 和 %s 为同一个人，相似度为：%f" % (person1, person2, dist))
+    # else:
+    #     print("%s 和 %s 不是同一个人，相似度为：%f" % (person1, person2, dist))
+
+    dirs = os.listdir('dataset/predict')
+    for dir in dirs:
+        dist_sum = 0
+        print(dir)
+        model_path = "models" + '/' + dir + '/resnet.h5' 
+        #model_path = os.path.join(model_path, 'resnet.h5')
+        print(model_path + " is running")
+        model = tf.keras.models.load_model(model_path)
+        intermediate_layer_model = Model(inputs=model.input, outputs=model.get_layer(layer_name).output)
+
+        person2 = 'dataset/wctest.wav'
+        feature2 = infer(person2)[0]
+
+        dir = os.path.join('dataset/predict', dir)
+        print(dir)
+        files = os.listdir(dir)
+        length = len(files)
+        predict_num = 20
+        for i in range(0, predict_num):
+            index = random.randint(0, length - 1)
+            # print(files[index])
+            person1 = os.path.join(dir, files[index])
+            feature1 = infer(person1)[0]
+            dist = np.dot(feature1, feature2) / (np.linalg.norm(feature1) * np.linalg.norm(feature2))
+            dist_sum = dist + dist_sum
+            if dist > 0.92:
+                print("%s 和 %s 为同一个人，相似度为：%f" % (person1, person2, dist))
+            else:
+                print("%s 和 %s 不是同一个人，相似度为：%f" % (person1, person2, dist))
+
+        dist_avg = dist_sum / predict_num
+        if dist_avg > 0.925:
+            print("为同一个人，相似度为：%f" % dist_avg)
+        else:
+            print("不是同一个人，相似度为：%f" % dist_avg)
+
+    # dist_sum = 0
+    # for file in files:
+    #     person1 = os.path.join('dataset/audio', file)
+    #     feature1 = infer(person1)[0]
+    #     dist = np.dot(feature1, feature2) / (np.linalg.norm(feature1) * np.linalg.norm(feature2))
+    #     dist_sum = dist + dist_sum
+    #     if dist > 0.92:
+    #         print("%s 和 %s 为同一个人，相似度为：%f" % (person1, person2, dist))
+    #     else:
+    #         print("%s 和 %s 不是同一个人，相似度为：%f" % (person1, person2, dist))
+    #
+    # num = len(files)
+    # dist_avg = dist_sum / num
+    # print(dist_avg)
+    # if dist_avg > 0.925:
+    #     print("为同一个人，相似度为：%f" % (dist_avg))
+    # else:
+    #     print("不是同一个人，相似度为：%f" % (dist_avg))

src/voiceprint/reader.py

@@ -0,0 +1,27 @@
+import tensorflow as tf
+
+
+def _parse_data_function(example):
+    # [可能需要修改参数】 设置的梅尔频谱的shape相乘的值
+    data_feature_description = {
+        'data': tf.io.FixedLenFeature([16384], tf.float32),
+        'label': tf.io.FixedLenFeature([], tf.int64),
+    }
+    return tf.io.parse_single_example(example, data_feature_description)
+
+
+def train_reader_tfrecord(data_path, num_epochs, batch_size):
+    raw_dataset = tf.data.TFRecordDataset(data_path)
+    train_dataset = raw_dataset.map(_parse_data_function)
+    train_dataset = train_dataset.shuffle(buffer_size=1000) \
+        .repeat(count=num_epochs) \
+        .batch(batch_size=batch_size) \
+        .prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
+    return train_dataset
+
+
+def test_reader_tfrecord(data_path, batch_size):
+    raw_dataset = tf.data.TFRecordDataset(data_path)
+    test_dataset = raw_dataset.map(_parse_data_function)
+    test_dataset = test_dataset.batch(batch_size=batch_size)
+    return test_dataset

src/voiceprint/train.py

@@ -0,0 +1,95 @@
+import os
+import sys
+
+import tensorflow as tf
+import reader
+import numpy as np
+
+class_dim = 1  # train_list里面有几个分类，这里的值要等于txt文件中的数字+1
+EPOCHS = 500
+BATCH_SIZE = 32
+init_model = None
+
+model = tf.keras.models.Sequential([
+    tf.keras.applications.ResNet50V2(include_top=False, weights=None, input_shape=(128, None, 1)),
+    tf.keras.layers.ActivityRegularization(l2=0.5),
+    tf.keras.layers.Dropout(rate=0.5),
+    tf.keras.layers.GlobalMaxPooling2D(),
+    tf.keras.layers.Dense(units=class_dim, activation=tf.nn.softmax)
+])
+
+model.summary()
+
+# 定义优化方法
+optimizer = tf.keras.optimizers.Adam(learning_rate=1e-3)
+
+workdir = sys.path[0]
+
+train_dataset = reader.train_reader_tfrecord(workdir + '/' + 'dataset/train.tfrecord', EPOCHS, batch_size=BATCH_SIZE)
+test_dataset = reader.test_reader_tfrecord(workdir + '/' + 'dataset/test.tfrecord', batch_size=BATCH_SIZE)
+
+if init_model:
+    model.load_weights(init_model)
+
+modelname = sys.argv[1]
+modeldir = workdir + '/' + 'models' + '/' + modelname
+print(modeldir)
+isExists = os.path.exists(modeldir)
+if not isExists:
+    os.makedirs(modeldir)
+    print(modeldir + ' 创建成功')
+else:
+    print(modeldir + ' 已存在')
+
+
+for batch_id, data in enumerate(train_dataset):
+    # [可能需要修改参数】 设置的梅尔频谱的shape
+    sounds = data['data'].numpy().reshape((-1, 128, 128, 1))
+    labels = data['label']
+    # 执行训练
+    with tf.GradientTape() as tape:
+        predictions = model(sounds)
+        # 获取损失值
+        train_loss = tf.keras.losses.sparse_categorical_crossentropy(labels, predictions)
+        train_loss = tf.reduce_mean(train_loss)
+        # 获取准确率
+        train_accuracy = tf.keras.metrics.sparse_categorical_accuracy(labels, predictions)
+        train_accuracy = np.sum(train_accuracy.numpy()) / len(train_accuracy.numpy())
+
+    # 更新梯度
+    gradients = tape.gradient(train_loss, model.trainable_variables)
+    optimizer.apply_gradients(zip(gradients, model.trainable_variables))
+
+    print(batch_id)
+    if batch_id % 20 == 0:
+        print("Batch %d, Loss %f, Accuracy %f" % (batch_id, train_loss.numpy(), train_accuracy))
+
+    if batch_id % 50 == 0 and batch_id != 0:
+        test_losses = list()
+        test_accuracies = list()
+        for d in test_dataset:
+            # [可能需要修改参数】 设置的梅尔频谱的shape
+            test_sounds = d['data'].numpy().reshape((-1, 128, 128, 1))
+            test_labels = d['label']
+
+            test_result = model(test_sounds)
+            # 获取损失值
+            test_loss = tf.keras.losses.sparse_categorical_crossentropy(test_labels, test_result)
+            test_loss = tf.reduce_mean(test_loss)
+            test_losses.append(test_loss)
+            # 获取准确率
+            test_accuracy = tf.keras.metrics.sparse_categorical_accuracy(test_labels, test_result)
+            test_accuracy = np.sum(test_accuracy.numpy()) / len(test_accuracy.numpy())
+            test_accuracies.append(test_accuracy)
+
+        print('=================================================')
+        print("Test, Loss %f, Accuracy %f" % (
+            sum(test_losses) / len(test_losses), sum(test_accuracies) / len(test_accuracies)))
+        print("models is creating")
+        print('=================================================')
+
+        # 保存模型
+        model.save(filepath=modeldir + '/resnet.h5')
+        model.save_weights(filepath=modeldir + '/model_weights.h5')
+
+print("训练完成")